aboutsummaryrefslogtreecommitdiff
path: root/llvm/lib/Target/AMDGPU
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/lib/Target/AMDGPU')
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPUGISel.td3
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp16
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.h3
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp10
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h2
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp6
-rw-r--r--llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp3
-rw-r--r--llvm/lib/Target/AMDGPU/FLATInstructions.td48
-rw-r--r--llvm/lib/Target/AMDGPU/GCNRegPressure.cpp6
-rw-r--r--llvm/lib/Target/AMDGPU/GCNRegPressure.h40
-rw-r--r--llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp11
-rw-r--r--llvm/lib/Target/AMDGPU/SIFoldOperands.cpp4
-rw-r--r--llvm/lib/Target/AMDGPU/SIISelLowering.cpp14
-rw-r--r--llvm/lib/Target/AMDGPU/SIInstrInfo.cpp57
-rw-r--r--llvm/lib/Target/AMDGPU/SIInstrInfo.h3
-rw-r--r--llvm/lib/Target/AMDGPU/SIInstrInfo.td1
-rw-r--r--llvm/lib/Target/AMDGPU/SIInstructions.td10
-rw-r--r--llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp14
-rw-r--r--llvm/lib/Target/AMDGPU/SIRegisterInfo.td43
-rw-r--r--llvm/lib/Target/AMDGPU/VOP3PInstructions.td6
20 files changed, 234 insertions, 66 deletions
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
index 108842f..c01e5d3 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
@@ -137,6 +137,9 @@ def gi_global_offset :
def gi_global_saddr :
GIComplexOperandMatcher<s64, "selectGlobalSAddr">,
GIComplexPatternEquiv<GlobalSAddr>;
+def gi_global_saddr_cpol :
+ GIComplexOperandMatcher<s64, "selectGlobalSAddrCPol">,
+ GIComplexPatternEquiv<GlobalSAddrCPol>;
def gi_global_saddr_glc :
GIComplexOperandMatcher<s64, "selectGlobalSAddrGLC">,
GIComplexPatternEquiv<GlobalSAddrGLC>;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index 0ca2286..dfaa145 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -2020,6 +2020,22 @@ bool AMDGPUDAGToDAGISel::SelectGlobalSAddr(SDNode *N, SDValue Addr,
return true;
}
+bool AMDGPUDAGToDAGISel::SelectGlobalSAddrCPol(SDNode *N, SDValue Addr,
+ SDValue &SAddr, SDValue &VOffset,
+ SDValue &Offset,
+ SDValue &CPol) const {
+ bool ScaleOffset;
+ if (!SelectGlobalSAddr(N, Addr, SAddr, VOffset, Offset, ScaleOffset))
+ return false;
+
+ // We are assuming CPol is always the last operand of the intrinsic.
+ auto PassedCPol =
+ N->getConstantOperandVal(N->getNumOperands() - 1) & ~AMDGPU::CPol::SCAL;
+ CPol = CurDAG->getTargetConstant(
+ (ScaleOffset ? AMDGPU::CPol::SCAL : 0) | PassedCPol, SDLoc(), MVT::i32);
+ return true;
+}
+
bool AMDGPUDAGToDAGISel::SelectGlobalSAddrGLC(SDNode *N, SDValue Addr,
SDValue &SAddr, SDValue &VOffset,
SDValue &Offset,
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.h b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.h
index a6ce745..5636d89 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.h
@@ -168,6 +168,9 @@ private:
bool SelectGlobalSAddr(SDNode *N, SDValue Addr, SDValue &SAddr,
SDValue &VOffset, SDValue &Offset,
SDValue &CPol) const;
+ bool SelectGlobalSAddrCPol(SDNode *N, SDValue Addr, SDValue &SAddr,
+ SDValue &VOffset, SDValue &Offset,
+ SDValue &CPol) const;
bool SelectGlobalSAddrGLC(SDNode *N, SDValue Addr, SDValue &SAddr,
SDValue &VOffset, SDValue &Offset,
SDValue &CPol) const;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index 8ca9a97..266dee1 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -5774,6 +5774,16 @@ AMDGPUInstructionSelector::selectGlobalSAddr(MachineOperand &Root) const {
}
InstructionSelector::ComplexRendererFns
+AMDGPUInstructionSelector::selectGlobalSAddrCPol(MachineOperand &Root) const {
+ const MachineInstr &I = *Root.getParent();
+
+ // We are assuming CPol is always the last operand of the intrinsic.
+ auto PassedCPol =
+ I.getOperand(I.getNumOperands() - 1).getImm() & ~AMDGPU::CPol::SCAL;
+ return selectGlobalSAddr(Root, PassedCPol);
+}
+
+InstructionSelector::ComplexRendererFns
AMDGPUInstructionSelector::selectGlobalSAddrGLC(MachineOperand &Root) const {
return selectGlobalSAddr(Root, AMDGPU::CPol::GLC);
}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h
index 61d9de1..fe9743d0a 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h
@@ -261,6 +261,8 @@ private:
InstructionSelector::ComplexRendererFns
selectGlobalSAddr(MachineOperand &Root) const;
InstructionSelector::ComplexRendererFns
+ selectGlobalSAddrCPol(MachineOperand &Root) const;
+ InstructionSelector::ComplexRendererFns
selectGlobalSAddrGLC(MachineOperand &Root) const;
InstructionSelector::ComplexRendererFns
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index 787db67..c5a1d9e 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -5180,6 +5180,12 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
case Intrinsic::amdgcn_ds_load_tr16_b128:
case Intrinsic::amdgcn_ds_load_tr4_b64:
case Intrinsic::amdgcn_ds_load_tr6_b96:
+ case Intrinsic::amdgcn_flat_load_monitor_b32:
+ case Intrinsic::amdgcn_flat_load_monitor_b64:
+ case Intrinsic::amdgcn_flat_load_monitor_b128:
+ case Intrinsic::amdgcn_global_load_monitor_b32:
+ case Intrinsic::amdgcn_global_load_monitor_b64:
+ case Intrinsic::amdgcn_global_load_monitor_b128:
case Intrinsic::amdgcn_ds_read_tr4_b64:
case Intrinsic::amdgcn_ds_read_tr6_b96:
case Intrinsic::amdgcn_ds_read_tr8_b64:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp b/llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp
index a8e1967..f580f43 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp
@@ -159,7 +159,8 @@ bool AMDGPURewriteAGPRCopyMFMAImpl::run(MachineFunction &MF) const {
// If the inputs are tied and the same register, we can shortcut and
// directly replace the register.
- if (Src2->getReg() != CopySrcReg) {
+ if (!Src2->isReg() || Src2->getReg() != CopySrcReg ||
+ Src2->getSubReg() != DefMI->getOperand(1).getSubReg()) {
LLVM_DEBUG(
dbgs()
<< "Replacing untied VGPR MFMAs with AGPR form not yet handled\n");
diff --git a/llvm/lib/Target/AMDGPU/FLATInstructions.td b/llvm/lib/Target/AMDGPU/FLATInstructions.td
index 5ccf1e5..7207c25 100644
--- a/llvm/lib/Target/AMDGPU/FLATInstructions.td
+++ b/llvm/lib/Target/AMDGPU/FLATInstructions.td
@@ -13,6 +13,7 @@ let WantsRoot = true in {
def GlobalSAddr : ComplexPattern<iPTR, 4, "SelectGlobalSAddr", [], [], -10>;
def GlobalSAddrGLC : ComplexPattern<iPTR, 4, "SelectGlobalSAddrGLC", [], [], -10>;
+ def GlobalSAddrCPol : ComplexPattern<iPTR, 4, "SelectGlobalSAddrCPol", [], [], -10>;
def ScratchSAddr : ComplexPattern<iPTR, 2, "SelectScratchSAddr", [], [], -10>;
def ScratchSVAddr : ComplexPattern<iPTR, 4, "SelectScratchSVAddr", [], [], -10>;
}
@@ -1274,6 +1275,11 @@ class FlatLoadPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCN
(inst $vaddr, $offset)
>;
+class FlatLoadPat_CPOL <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
+ (vt (node (FlatOffset i64:$vaddr, i32:$offset), (i32 timm:$cpol))),
+ (inst $vaddr, $offset, $cpol)
+>;
+
class FlatLoadPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
(node (FlatOffset (i64 VReg_64:$vaddr), i32:$offset), vt:$in),
(inst $vaddr, $offset, 0, $in)
@@ -1295,8 +1301,8 @@ class FlatSignedLoadPat_D16_t16 <FLAT_Pseudo inst, SDPatternOperator node, Value
>;
class GlobalLoadSaddrPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
- (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i32:$offset), vt:$in)),
- (inst $saddr, $voffset, $offset, 0, $in)
+ (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i32:$offset, CPol:$cpol), vt:$in)),
+ (inst $saddr, $voffset, $offset, $cpol, $in)
>;
class FlatLoadSaddrPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
@@ -1310,8 +1316,8 @@ class FlatLoadSaddrPat_D16_t16 <FLAT_Pseudo inst, SDPatternOperator node, ValueT
>;
class GlobalLoadSaddrPat_D16_t16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
- (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i32:$offset))),
- (inst $saddr, $voffset, $offset, (i32 0))
+ (vt (node (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i32:$offset, CPol:$cpol))),
+ (inst $saddr, $voffset, $offset, $cpol)
>;
class FlatLoadSignedPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
@@ -1324,6 +1330,16 @@ class FlatLoadSaddrPat <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt>
(inst $saddr, $voffset, $offset, $cpol)
>;
+class FlatLoadSignedPat_CPOL <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
+ (vt (node (GlobalOffset (i64 VReg_64:$vaddr), i32:$offset), (i32 timm:$cpol))),
+ (inst $vaddr, $offset, $cpol)
+>;
+
+class GlobalLoadSaddrPat_CPOL <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
+ (vt (node (GlobalSAddrCPol (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i32:$offset, CPol:$cpol), (i32 timm))),
+ (inst $saddr, $voffset, $offset, $cpol)
+>;
+
class FlatStoreSaddrPat <FLAT_Pseudo inst, SDPatternOperator node,
ValueType vt> : GCNPat <
(node vt:$data, (GlobalSAddr (i64 SReg_64:$saddr), (i32 VGPR_32:$voffset), i32:$offset, CPol:$cpol)),
@@ -1505,8 +1521,8 @@ class ScratchLoadSVaddrPat_D16 <FLAT_Pseudo inst, SDPatternOperator node, ValueT
>;
class ScratchLoadSVaddrPat_D16_t16 <FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> : GCNPat <
- (vt (node (ScratchSVAddr (i32 VGPR_32:$vaddr), (i32 SGPR_32:$saddr), i32:$offset))),
- (inst $vaddr, $saddr, $offset, 0)
+ (vt (node (ScratchSVAddr (i32 VGPR_32:$vaddr), (i32 SGPR_32:$saddr), i32:$offset, CPol:$cpol))),
+ (inst $vaddr, $saddr, $offset, $cpol)
>;
multiclass GlobalFLATLoadPats<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
@@ -1519,6 +1535,16 @@ multiclass GlobalFLATLoadPats<FLAT_Pseudo inst, SDPatternOperator node, ValueTyp
}
}
+multiclass GlobalFLATLoadPats_CPOL<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
+ def : FlatLoadSignedPat_CPOL<inst, node, vt> {
+ let AddedComplexity = 10;
+ }
+
+ def : GlobalLoadSaddrPat_CPOL<!cast<FLAT_Pseudo>(!cast<string>(inst)#"_SADDR"), node, vt> {
+ let AddedComplexity = 11;
+ }
+}
+
multiclass GlobalFLATLoadPats_D16<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
def : FlatSignedLoadPat_D16 <inst, node, vt> {
let AddedComplexity = 10;
@@ -2055,6 +2081,16 @@ let WaveSizePredicate = isWave32, OtherPredicates = [HasTransposeLoadF4F6Insts]
defm : GlobalFLATLoadPats <GLOBAL_LOAD_TR6_B96, int_amdgcn_global_load_tr6_b96, v3i32>;
}
+let OtherPredicates = [isGFX125xOnly] in {
+ def : FlatLoadPat_CPOL <FLAT_LOAD_MONITOR_B32, int_amdgcn_flat_load_monitor_b32, i32>;
+ def : FlatLoadPat_CPOL <FLAT_LOAD_MONITOR_B64, int_amdgcn_flat_load_monitor_b64, v2i32>;
+ def : FlatLoadPat_CPOL <FLAT_LOAD_MONITOR_B128, int_amdgcn_flat_load_monitor_b128, v4i32>;
+
+ defm : GlobalFLATLoadPats_CPOL <GLOBAL_LOAD_MONITOR_B32, int_amdgcn_global_load_monitor_b32, i32>;
+ defm : GlobalFLATLoadPats_CPOL <GLOBAL_LOAD_MONITOR_B64, int_amdgcn_global_load_monitor_b64, v2i32>;
+ defm : GlobalFLATLoadPats_CPOL <GLOBAL_LOAD_MONITOR_B128, int_amdgcn_global_load_monitor_b128, v4i32>;
+} // End SubtargetPredicate = isGFX125xOnly
+
let SubtargetPredicate = HasAtomicFMinFMaxF32GlobalInsts, OtherPredicates = [HasFlatGlobalInsts] in {
defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_FMIN", "atomic_load_fmin_global", f32>;
defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_FMAX", "atomic_load_fmax_global", f32>;
diff --git a/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp b/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
index 7d6723a..334afd3 100644
--- a/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
@@ -38,7 +38,11 @@ bool llvm::isEqual(const GCNRPTracker::LiveRegSet &S1,
unsigned GCNRegPressure::getRegKind(const TargetRegisterClass *RC,
const SIRegisterInfo *STI) {
- return STI->isSGPRClass(RC) ? SGPR : (STI->isAGPRClass(RC) ? AGPR : VGPR);
+ return STI->isSGPRClass(RC)
+ ? SGPR
+ : (STI->isAGPRClass(RC)
+ ? AGPR
+ : (STI->isVectorSuperClass(RC) ? AVGPR : VGPR));
}
void GCNRegPressure::inc(unsigned Reg,
diff --git a/llvm/lib/Target/AMDGPU/GCNRegPressure.h b/llvm/lib/Target/AMDGPU/GCNRegPressure.h
index 3749b6d..ea33a22 100644
--- a/llvm/lib/Target/AMDGPU/GCNRegPressure.h
+++ b/llvm/lib/Target/AMDGPU/GCNRegPressure.h
@@ -29,43 +29,57 @@ class raw_ostream;
class SlotIndex;
struct GCNRegPressure {
- enum RegKind { SGPR, VGPR, AGPR, TOTAL_KINDS };
+ enum RegKind { SGPR, VGPR, AGPR, AVGPR, TOTAL_KINDS };
GCNRegPressure() {
clear();
}
- bool empty() const { return !Value[SGPR] && !Value[VGPR] && !Value[AGPR]; }
+ bool empty() const {
+ return !Value[SGPR] && !Value[VGPR] && !Value[AGPR] && !Value[AVGPR];
+ }
void clear() { std::fill(&Value[0], &Value[ValueArraySize], 0); }
/// \returns the SGPR32 pressure
unsigned getSGPRNum() const { return Value[SGPR]; }
- /// \returns the aggregated ArchVGPR32, AccVGPR32 pressure dependent upon \p
- /// UnifiedVGPRFile
+ /// \returns the aggregated ArchVGPR32, AccVGPR32, and Pseudo AVGPR pressure
+ /// dependent upon \p UnifiedVGPRFile
unsigned getVGPRNum(bool UnifiedVGPRFile) const {
if (UnifiedVGPRFile) {
- return Value[AGPR] ? getUnifiedVGPRNum(Value[VGPR], Value[AGPR])
- : Value[VGPR];
+ return Value[AGPR]
+ ? getUnifiedVGPRNum(Value[VGPR], Value[AGPR], Value[AVGPR])
+ : Value[VGPR] + Value[AVGPR];
}
- return std::max(Value[VGPR], Value[AGPR]);
+ // AVGPR assignment priority is based on the width of the register. Account
+ // AVGPR pressure as VGPR.
+ return std::max(Value[VGPR] + Value[AVGPR], Value[AGPR]);
}
/// Returns the aggregated VGPR pressure, assuming \p NumArchVGPRs ArchVGPRs
- /// and \p NumAGPRs AGPRS, for a target with a unified VGPR file.
+ /// \p NumAGPRs AGPRS, and \p NumAVGPRs AVGPRs for a target with a unified
+ /// VGPR file.
inline static unsigned getUnifiedVGPRNum(unsigned NumArchVGPRs,
- unsigned NumAGPRs) {
- return alignTo(NumArchVGPRs, AMDGPU::IsaInfo::getArchVGPRAllocGranule()) +
+ unsigned NumAGPRs,
+ unsigned NumAVGPRs) {
+
+ // Assume AVGPRs will be assigned as VGPRs.
+ return alignTo(NumArchVGPRs + NumAVGPRs,
+ AMDGPU::IsaInfo::getArchVGPRAllocGranule()) +
NumAGPRs;
}
- /// \returns the ArchVGPR32 pressure
- unsigned getArchVGPRNum() const { return Value[VGPR]; }
+ /// \returns the ArchVGPR32 pressure, plus the AVGPRS which we assume will be
+ /// allocated as VGPR
+ unsigned getArchVGPRNum() const { return Value[VGPR] + Value[AVGPR]; }
/// \returns the AccVGPR32 pressure
unsigned getAGPRNum() const { return Value[AGPR]; }
+ /// \returns the AVGPR32 pressure
+ unsigned getAVGPRNum() const { return Value[AVGPR]; }
unsigned getVGPRTuplesWeight() const {
- return std::max(Value[TOTAL_KINDS + VGPR], Value[TOTAL_KINDS + AGPR]);
+ return std::max(Value[TOTAL_KINDS + VGPR] + Value[TOTAL_KINDS + AVGPR],
+ Value[TOTAL_KINDS + AGPR]);
}
unsigned getSGPRTuplesWeight() const { return Value[TOTAL_KINDS + SGPR]; }
diff --git a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
index a655308..ce1ce68 100644
--- a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
@@ -1911,14 +1911,12 @@ void PreRARematStage::rematerialize() {
for (auto &[DefMI, Remat] : Rematerializations) {
MachineBasicBlock::iterator InsertPos(Remat.UseMI);
Register Reg = DefMI->getOperand(0).getReg();
- unsigned SubReg = DefMI->getOperand(0).getSubReg();
unsigned DefRegion = MIRegion.at(DefMI);
// Rematerialize DefMI to its use block.
- TII->reMaterialize(*InsertPos->getParent(), InsertPos, Reg, SubReg, *DefMI,
- *DAG.TRI);
+ TII->reMaterialize(*InsertPos->getParent(), InsertPos, Reg,
+ AMDGPU::NoSubRegister, *DefMI, *DAG.TRI);
Remat.RematMI = &*std::prev(InsertPos);
- Remat.RematMI->getOperand(0).setSubReg(SubReg);
DAG.LIS->InsertMachineInstrInMaps(*Remat.RematMI);
// Update region boundaries in regions we sinked from (remove defining MI)
@@ -2064,14 +2062,13 @@ void PreRARematStage::finalizeGCNSchedStage() {
MachineBasicBlock::iterator InsertPos(DAG.Regions[DefRegion].second);
MachineBasicBlock *MBB = RegionBB[DefRegion];
Register Reg = RematMI.getOperand(0).getReg();
- unsigned SubReg = RematMI.getOperand(0).getSubReg();
// Re-rematerialize MI at the end of its original region. Note that it may
// not be rematerialized exactly in the same position as originally within
// the region, but it should not matter much.
- TII->reMaterialize(*MBB, InsertPos, Reg, SubReg, RematMI, *DAG.TRI);
+ TII->reMaterialize(*MBB, InsertPos, Reg, AMDGPU::NoSubRegister, RematMI,
+ *DAG.TRI);
MachineInstr *NewMI = &*std::prev(InsertPos);
- NewMI->getOperand(0).setSubReg(SubReg);
DAG.LIS->InsertMachineInstrInMaps(*NewMI);
auto UseRegion = MIRegion.find(Remat.UseMI);
diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
index e5d1eaa..b77da4d 100644
--- a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -1062,9 +1062,13 @@ bool SIFoldOperandsImpl::tryFoldRegSeqSplat(
switch (OpTy) {
case AMDGPU::OPERAND_REG_INLINE_AC_INT32:
case AMDGPU::OPERAND_REG_INLINE_AC_FP32:
+ case AMDGPU::OPERAND_REG_INLINE_C_INT32:
+ case AMDGPU::OPERAND_REG_INLINE_C_FP32:
OpRC = TRI->getSubRegisterClass(OpRC, AMDGPU::sub0);
break;
case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
+ case AMDGPU::OPERAND_REG_INLINE_C_FP64:
+ case AMDGPU::OPERAND_REG_INLINE_C_INT64:
OpRC = TRI->getSubRegisterClass(OpRC, AMDGPU::sub0_sub1);
break;
default:
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 74fe2b8..8d51ec6 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1477,6 +1477,12 @@ bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
MachineMemOperand::MOVolatile;
return true;
}
+ case Intrinsic::amdgcn_flat_load_monitor_b32:
+ case Intrinsic::amdgcn_flat_load_monitor_b64:
+ case Intrinsic::amdgcn_flat_load_monitor_b128:
+ case Intrinsic::amdgcn_global_load_monitor_b32:
+ case Intrinsic::amdgcn_global_load_monitor_b64:
+ case Intrinsic::amdgcn_global_load_monitor_b128:
case Intrinsic::amdgcn_ds_load_tr6_b96:
case Intrinsic::amdgcn_ds_load_tr4_b64:
case Intrinsic::amdgcn_ds_load_tr8_b64:
@@ -1603,10 +1609,16 @@ bool SITargetLowering::getAddrModeArguments(const IntrinsicInst *II,
case Intrinsic::amdgcn_ds_atomic_barrier_arrive_rtn_b64:
case Intrinsic::amdgcn_flat_atomic_fmax_num:
case Intrinsic::amdgcn_flat_atomic_fmin_num:
+ case Intrinsic::amdgcn_flat_load_monitor_b128:
+ case Intrinsic::amdgcn_flat_load_monitor_b32:
+ case Intrinsic::amdgcn_flat_load_monitor_b64:
case Intrinsic::amdgcn_global_atomic_csub:
case Intrinsic::amdgcn_global_atomic_fmax_num:
case Intrinsic::amdgcn_global_atomic_fmin_num:
case Intrinsic::amdgcn_global_atomic_ordered_add_b64:
+ case Intrinsic::amdgcn_global_load_monitor_b128:
+ case Intrinsic::amdgcn_global_load_monitor_b32:
+ case Intrinsic::amdgcn_global_load_monitor_b64:
case Intrinsic::amdgcn_global_load_tr_b64:
case Intrinsic::amdgcn_global_load_tr_b128:
case Intrinsic::amdgcn_global_load_tr4_b64:
@@ -14167,6 +14179,8 @@ SDValue SITargetLowering::performMinMaxCombine(SDNode *N,
Op0.getOpcode() == AMDGPUISD::FMAX_LEGACY)) &&
(VT == MVT::f32 || VT == MVT::f64 ||
(VT == MVT::f16 && Subtarget->has16BitInsts()) ||
+ (VT == MVT::bf16 && Subtarget->hasBF16PackedInsts()) ||
+ (VT == MVT::v2bf16 && Subtarget->hasBF16PackedInsts()) ||
(VT == MVT::v2f16 && Subtarget->hasVOP3PInsts())) &&
Op0.hasOneUse()) {
if (SDValue Res = performFPMed3ImmCombine(DAG, SDLoc(N), Op0, Op1))
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 40e6871..8d6c1d0 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2508,7 +2508,20 @@ bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
.addReg(DstHi);
}
break;
+
+ case AMDGPU::V_MAX_BF16_PSEUDO_e64:
+ assert(ST.hasBF16PackedInsts());
+ MI.setDesc(get(AMDGPU::V_PK_MAX_NUM_BF16));
+ MI.addOperand(MachineOperand::CreateImm(0)); // op_sel
+ MI.addOperand(MachineOperand::CreateImm(0)); // neg_lo
+ MI.addOperand(MachineOperand::CreateImm(0)); // neg_hi
+ auto Op0 = getNamedOperand(MI, AMDGPU::OpName::src0_modifiers);
+ Op0->setImm(Op0->getImm() | SISrcMods::OP_SEL_1);
+ auto Op1 = getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);
+ Op1->setImm(Op1->getImm() | SISrcMods::OP_SEL_1);
+ break;
}
+
return true;
}
@@ -2733,49 +2746,47 @@ static MachineInstr *swapImmOperands(MachineInstr &MI,
}
bool SIInstrInfo::isLegalToSwap(const MachineInstr &MI, unsigned OpIdx0,
- const MachineOperand *MO0, unsigned OpIdx1,
- const MachineOperand *MO1) const {
+ unsigned OpIdx1) const {
const MCInstrDesc &InstDesc = MI.getDesc();
const MCOperandInfo &OpInfo0 = InstDesc.operands()[OpIdx0];
const MCOperandInfo &OpInfo1 = InstDesc.operands()[OpIdx1];
- const TargetRegisterClass *DefinedRC1 =
- OpInfo1.RegClass != -1 ? RI.getRegClass(OpInfo1.RegClass) : nullptr;
- const TargetRegisterClass *DefinedRC0 =
- OpInfo1.RegClass != -1 ? RI.getRegClass(OpInfo0.RegClass) : nullptr;
unsigned Opc = MI.getOpcode();
int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
+ const MachineOperand &MO0 = MI.getOperand(OpIdx0);
+ const MachineOperand &MO1 = MI.getOperand(OpIdx1);
+
// Swap doesn't breach constant bus or literal limits
// It may move literal to position other than src0, this is not allowed
// pre-gfx10 However, most test cases need literals in Src0 for VOP
// FIXME: After gfx9, literal can be in place other than Src0
if (isVALU(MI)) {
- if ((int)OpIdx0 == Src0Idx && !MO0->isReg() &&
- !isInlineConstant(*MO0, OpInfo1))
+ if ((int)OpIdx0 == Src0Idx && !MO0.isReg() &&
+ !isInlineConstant(MO0, OpInfo1))
return false;
- if ((int)OpIdx1 == Src0Idx && !MO1->isReg() &&
- !isInlineConstant(*MO1, OpInfo0))
+ if ((int)OpIdx1 == Src0Idx && !MO1.isReg() &&
+ !isInlineConstant(MO1, OpInfo0))
return false;
}
- if ((int)OpIdx1 != Src0Idx && MO0->isReg()) {
- if (!DefinedRC1)
+ if ((int)OpIdx1 != Src0Idx && MO0.isReg()) {
+ if (OpInfo1.RegClass == -1)
return OpInfo1.OperandType == MCOI::OPERAND_UNKNOWN;
- return isLegalRegOperand(MI, OpIdx1, *MO0) &&
- (!MO1->isReg() || isLegalRegOperand(MI, OpIdx0, *MO1));
+ return isLegalRegOperand(MI, OpIdx1, MO0) &&
+ (!MO1.isReg() || isLegalRegOperand(MI, OpIdx0, MO1));
}
- if ((int)OpIdx0 != Src0Idx && MO1->isReg()) {
- if (!DefinedRC0)
+ if ((int)OpIdx0 != Src0Idx && MO1.isReg()) {
+ if (OpInfo0.RegClass == -1)
return OpInfo0.OperandType == MCOI::OPERAND_UNKNOWN;
- return (!MO0->isReg() || isLegalRegOperand(MI, OpIdx1, *MO0)) &&
- isLegalRegOperand(MI, OpIdx0, *MO1);
+ return (!MO0.isReg() || isLegalRegOperand(MI, OpIdx1, MO0)) &&
+ isLegalRegOperand(MI, OpIdx0, MO1);
}
// No need to check 64-bit literals since swapping does not bring new
// 64-bit literals into current instruction to fold to 32-bit
- return isImmOperandLegal(MI, OpIdx1, *MO0);
+ return isImmOperandLegal(MI, OpIdx1, MO0);
}
MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
@@ -2797,12 +2808,12 @@ MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
static_cast<int>(Src1Idx) &&
"inconsistency with findCommutedOpIndices");
- MachineOperand &Src0 = MI.getOperand(Src0Idx);
- MachineOperand &Src1 = MI.getOperand(Src1Idx);
- if (!isLegalToSwap(MI, Src0Idx, &Src0, Src1Idx, &Src1)) {
+ if (!isLegalToSwap(MI, Src0Idx, Src1Idx))
return nullptr;
- }
+
MachineInstr *CommutedMI = nullptr;
+ MachineOperand &Src0 = MI.getOperand(Src0Idx);
+ MachineOperand &Src1 = MI.getOperand(Src1Idx);
if (Src0.isReg() && Src1.isReg()) {
// Be sure to copy the source modifiers to the right place.
CommutedMI =
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
index 800ea9a..2ffb783 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -197,8 +197,7 @@ protected:
AMDGPU::OpName Src0OpName, MachineOperand &Src1,
AMDGPU::OpName Src1OpName) const;
bool isLegalToSwap(const MachineInstr &MI, unsigned fromIdx,
- const MachineOperand *fromMO, unsigned toIdx,
- const MachineOperand *toMO) const;
+ unsigned toIdx) const;
MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
unsigned OpIdx0,
unsigned OpIdx1) const override;
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index b0be3f86..83b0490 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -2865,6 +2865,7 @@ def VOP_F16_F16_I16 : VOPProfile <[f16, f16, i16, untyped]>;
def VOP_F16_F16_I32 : VOPProfile <[f16, f16, i32, untyped]>;
def VOP_I16_I16_I16 : VOPProfile <[i16, i16, i16, untyped]>;
def VOP_I16_I16_I16_ARITH : VOPProfile <[i16, i16, i16, untyped], /*EnableClamp=*/1>;
+def VOP_BF16_BF16_BF16 : VOPProfile <[bf16, bf16, bf16, untyped]>;
def VOP_I16_I16_I16_I16 : VOPProfile <[i16, i16, i16, i16, untyped]>;
def VOP_F16_F16_F16_F16 : VOPProfile <[f16, f16, f16, f16, untyped]>;
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index d05be8f..54fa192 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -1894,6 +1894,9 @@ let SubtargetPredicate = UseRealTrue16Insts in
def : ClampPat<V_MAX_F16_t16_e64, f16>;
let SubtargetPredicate = UseFakeTrue16Insts in
def : ClampPat<V_MAX_F16_fake16_e64, f16>;
+// FIXME-TRUE16: Pseudo expansion of this won't work with True16.
+let True16Predicate = UseFakeTrue16Insts in
+def : ClampPat<V_MAX_BF16_PSEUDO_e64, bf16>;
let SubtargetPredicate = HasVOP3PInsts in {
def : GCNPat <
@@ -1903,6 +1906,13 @@ def : GCNPat <
>;
}
+let SubtargetPredicate = HasBF16PackedInsts in {
+def : GCNPat <
+ (v2bf16 (AMDGPUclamp (VOP3PMods v2bf16:$src0, i32:$src0_modifiers))),
+ (V_PK_MAX_NUM_BF16 $src0_modifiers, $src0,
+ $src0_modifiers, $src0, DSTCLAMP.ENABLE)
+>;
+} // End SubtargetPredicate = HasBF16PackedInsts
/********** ================================ **********/
/********** Floating point absolute/negative **********/
diff --git a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
index f0be204..9a1448f 100644
--- a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
@@ -81,11 +81,15 @@ SIMachineFunctionInfo::SIMachineFunctionInfo(const Function &F,
PSInputAddr = AMDGPU::getInitialPSInputAddr(F);
}
- MayNeedAGPRs = ST.hasMAIInsts() && !MFMAVGPRForm;
- if (!MFMAVGPRForm && ST.hasGFX90AInsts() &&
- ST.getMaxNumVGPRs(F) <= AMDGPU::VGPR_32RegClass.getNumRegs() &&
- !mayUseAGPRs(F))
- MayNeedAGPRs = false; // We will select all MAI with VGPR operands.
+ MayNeedAGPRs = ST.hasMAIInsts();
+ if (ST.hasGFX90AInsts()) {
+ // FIXME: MayNeedAGPRs is a misnomer for how this is used. MFMA selection
+ // should be separated from availability of AGPRs
+ if (MFMAVGPRForm ||
+ (ST.getMaxNumVGPRs(F) <= AMDGPU::VGPR_32RegClass.getNumRegs() &&
+ !mayUseAGPRs(F)))
+ MayNeedAGPRs = false; // We will select all MAI with VGPR operands.
+ }
if (AMDGPU::isChainCC(CC)) {
// Chain functions don't receive an SP from their caller, but are free to
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
index 0039d2f..218841d 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
@@ -109,6 +109,23 @@ class SIRegisterClass <string n, list<ValueType> rTypes, int Align, dag rList>
let TSFlags{2} = HasVGPR;
let TSFlags{3} = HasAGPR;
let TSFlags{4} = HasSGPR;
+
+ // RA will use RegisterClass AllocationPriority amongst other info (e.g. ordering in the basic block)
+ // to decide which registers to try to assign first. Usually, this RegisterClass priority is given
+ // very high priority, if not the highest priority, when considering which VirtReg to allocate next.
+ //
+ // We have 5 bits to assign AllocationPriorities to RegisterClasses. Generally, it is beneficial to
+ // assign more constrained RegisterClasses first. As a result, we prioritize register classes with
+ // more 32 bit tuples (e.g. VReg_512) over registers with fewer tuples (e.g. VGPR_32).
+ //
+ // The interesting case is the vector register case on architectures which have ARegs, VRegs, AVRegs.
+ // In this case, we would like to assign ARegs and VRegs before AVRegs, as AVRegs are less constrained
+ // and can be assigned to both AGPRs and VGPRs. We use the 5th bit to encode this into the
+ // RegisterClass AllocationPriority. BaseClassPriority is used to turn the bit on, and BaseClassScaleFactor
+ // is used for scaling of the bit (i.e. 1 << 4).
+ field int BaseClassPriority = 1;
+ field int BaseClassScaleFactor = 16;
+
}
multiclass SIRegLoHi16 <string n, bits<8> regIdx, bit ArtificialHigh = 1,
@@ -575,7 +592,7 @@ let HasVGPR = 1 in {
def VGPR_16 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
(add (interleave (sequence "VGPR%u_LO16", 0, 255),
(sequence "VGPR%u_HI16", 0, 255)))> {
- let AllocationPriority = 2;
+ let AllocationPriority = !add(2, !mul(BaseClassPriority, BaseClassScaleFactor));
let Size = 16;
let GeneratePressureSet = 0;
@@ -601,7 +618,7 @@ def VGPR_16_Lo128 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
// i16/f16 only on VI+
def VGPR_32 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types.types), 32,
(add (sequence "VGPR%u", 0, 255))> {
- let AllocationPriority = 0;
+ let AllocationPriority = !add(0, !mul(BaseClassPriority, BaseClassScaleFactor));
let Size = 32;
let Weight = 1;
let BaseClassOrder = 32;
@@ -610,7 +627,7 @@ def VGPR_32 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types
// Identical to VGPR_32 except it only contains the low 128 (Lo128) registers.
def VGPR_32_Lo128 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types.types), 32,
(add (sequence "VGPR%u", 0, 127))> {
- let AllocationPriority = 0;
+ let AllocationPriority = !add(0, !mul(BaseClassPriority, BaseClassScaleFactor));
let GeneratePressureSet = 0;
let Size = 32;
let Weight = 1;
@@ -668,7 +685,7 @@ def AGPR_LO16 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
// AccVGPR 32-bit registers
def AGPR_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, bf16, v2i16, v2f16, v2bf16], 32,
(add (sequence "AGPR%u", 0, 255))> {
- let AllocationPriority = 0;
+ let AllocationPriority = !add(0, !mul(BaseClassPriority, BaseClassScaleFactor));
let Size = 32;
let Weight = 1;
let BaseClassOrder = 32;
@@ -940,14 +957,23 @@ class VRegClassBase<int numRegs, list<ValueType> regTypes, dag regList> :
// Requires n v_mov_b32 to copy
let CopyCost = numRegs;
- let AllocationPriority = !sub(numRegs, 1);
+
+ // Since we only have 5 bits for the RegisterClass Allocation Priorty, and since we use the
+ // 5th bit for BaseClassPriority, we need to encode the SizePriority into 4 bits. As a result
+ // of this encoding, for registers with numRegs 15 or 16, we give SizePriority of 14, and for
+ // regsters with numRegs 17+ we give SizePriority of 15. In practice, there is only one
+ // RegClass per Vector Register type in each of these groups (i.e. numRegs = 15,16 : {VReg_512},
+ // and numRegs = 17+ : {VReg_1024}). Therefore, we have not lost any info by compressing.
+ defvar SizePrioriity = !if(!le(numRegs, 14), !sub(numRegs, 1), !if(!le(numRegs, 16), 14, 15));
+
+ let AllocationPriority = !add(SizePrioriity, !mul(BaseClassPriority, BaseClassScaleFactor));
let Weight = numRegs;
}
// Define a register tuple class, along with one requiring an even
// aligned base register.
multiclass VRegClass<int numRegs, list<ValueType> regTypes, dag regList> {
- let HasVGPR = 1 in {
+ let HasVGPR = 1, BaseClassPriority = 1 in {
// Define the regular class.
def "" : VRegClassBase<numRegs, regTypes, regList> {
let BaseClassOrder = !mul(numRegs, 32);
@@ -981,7 +1007,7 @@ defm VReg_1024 : VRegClass<32, Reg1024Types.types, (add VGPR_1024)>;
}
multiclass ARegClass<int numRegs, list<ValueType> regTypes, dag regList> {
- let CopyCost = !add(numRegs, numRegs, 1), HasAGPR = 1 in {
+ let CopyCost = !add(numRegs, numRegs, 1), HasAGPR = 1, BaseClassPriority = 1 in {
// Define the regular class.
def "" : VRegClassBase<numRegs, regTypes, regList> {
let BaseClassOrder = !mul(numRegs, 32);
@@ -1066,6 +1092,7 @@ def VS_64 : SIRegisterClass<"AMDGPU", VReg_64.RegTypes, 32, (add VReg_64, SReg_6
def AV_32 : SIRegisterClass<"AMDGPU", VGPR_32.RegTypes, 32, (add VGPR_32, AGPR_32)> {
let HasVGPR = 1;
let HasAGPR = 1;
+ let BaseClassPriority = 0;
let Size = 32;
}
} // End GeneratePressureSet = 0
@@ -1074,7 +1101,7 @@ def AV_32 : SIRegisterClass<"AMDGPU", VGPR_32.RegTypes, 32, (add VGPR_32, AGPR_3
// aligned base register.
multiclass AVRegClass<int numRegs, list<ValueType> regTypes,
dag vregList, dag aregList> {
- let HasVGPR = 1, HasAGPR = 1 in {
+ let HasVGPR = 1, HasAGPR = 1, BaseClassPriority = 0 in {
// Define the regular class.
def "" : VRegClassBase<numRegs, regTypes, (add vregList, aregList)>;
diff --git a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
index c812dc9..95fcd4a 100644
--- a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
@@ -1236,6 +1236,12 @@ let isCommutable = 1, isReMaterializable = 1 in {
defm V_PK_MIN_NUM_BF16 : VOP3PInst<"v_pk_min_num_bf16", VOP3P_Profile<VOP_V2BF16_V2BF16_V2BF16, VOP3_PACKED>, fminnum_like>;
defm V_PK_MAX_NUM_BF16 : VOP3PInst<"v_pk_max_num_bf16", VOP3P_Profile<VOP_V2BF16_V2BF16_V2BF16, VOP3_PACKED>, fmaxnum_like>;
defm V_PK_FMA_BF16 : VOP3PInst<"v_pk_fma_bf16", VOP3P_Profile<VOP_V2BF16_V2BF16_V2BF16_V2BF16, VOP3_PACKED>, any_fma>;
+
+ // Scalar pseudo used to emulate AMDGPUClamp.
+ // Expanded to V_PK_MAX_NUM_BF16 with unused high half.
+ // FIXME-TRUE16: Pseudo expansion of this won't work with True16.
+ let True16Predicate = UseFakeTrue16Insts in
+ defm V_MAX_BF16_PSEUDO : VOP3Inst <"v_max_bf16", VOP_BF16_BF16_BF16>;
}
} // End isCommutable = 1, isReMaterializable = 1